Anthropic’s IPO filing lists $518 billion in commitments as clients try cheaper AI

Anthropic‘s pitch to public-market investors now has a paper trail, and its most revealing line is not about superintelligence. The draft prospectus for the Claude maker’s stock-market listing, reviewed by Reuters, warns that many of its largest customers are not bound by long-term contracts and could reduce or stop their spending. That caveat surfaces just as corporate buyers are showing how easily routine AI work can move to far cheaper models.

The collision matters outside Silicon Valley because the AI boom is financed on the assumption that businesses will keep paying premium prices for the most capable models. If dating apps, telecom operators and banks decide most everyday tasks can run on software that costs a fraction as much, the bill lands on the labs, on the cloud providers building their data centers, and eventually on the pension funds and savers asked to buy the shares.

The filing shows both the momentum and the exposure. Revenue grew roughly twelvefold last year to nearly $4.6 billion, while the operating loss widened to about $8 billion from about $3 billion. The $42 billion net loss overstates the damage, since about $34 billion of it is an accounting charge on financing that can convert into shares. Compute and infrastructure cost $7.33 billion, more than half of operating expenses, and the company lists $518 billion in future cloud and infrastructure obligations against roughly $20 billion in cash and short-term investments.

Two unnamed customers accounted for nearly a quarter of revenue. That concentration is the pressure point, because buyers are getting more price-sensitive. The Financial Times reported that mentions of open-weight models on earnings calls rose sixfold from a year earlier, with PNC Financial, CH Robinson and Siemens among the adopters. Tinder’s chief technology officer, Vinay Kuruvila, told the paper the app’s AI spending climbed from an annual rate of about $1 million in January to $10 million by July, and that he did not want another tenfold jump; Tinder now routes routine requests to open-weight models. AT&T executive Mark Austin has said about 40% of the company’s internal AI requests already go to open models, with a target of 60% to 70%.

Usage data agrees. On Vercel’s AI Gateway, which developers use to route requests between models, open-weight systems handled 56% of tokens in August, up from 7% in December, with Chinese-built families such as DeepSeek, Alibaba’s Qwen and Z.ai’s GLM behind much of the growth. The average price paid per token on the gateway fell 23.2% in August, the third monthly decline in a row.

Volume, however, is not revenue. The same Vercel figures show open models took only about 14% of spending while carrying more than half the traffic, and Anthropic alone captured 64% of all gateway spend in August. Cheap models are winning bulk work such as summaries and customer-service chat, while companies keep paying for frontier systems where a mistake is expensive; even AT&T still sends complex tasks to OpenAI and Anthropic. Router data also over-represents developers who switch at the first cheaper option, and it misses the enterprise contracts and direct API traffic that account for much of the labs’ income.

Those most exposed to that distinction are the ones asked to underwrite it. Reuters reported that Anthropic is targeting a valuation above $2 trillion, more than double the $965 billion set in its last private round, in effect a bet that the premium tier holds as the cheap tier grows. The cloud providers and chipmakers on the other side of the $518 billion in commitments carry a version of the same bet. The prospectus devotes roughly a third of its 261 pages to risk, Fortune reported, including warnings that AI could pose catastrophic or existential risks.

Anthropic filed confidentially for its listing on June 1 and has not made the document public; Reuters said the offering is expected after November. A harder read on AI budgets comes sooner, when Micron reports on September 30 against guidance of about $50 billion in quarterly revenue. Until then, investors are pricing a company whose biggest clients can leave, in a market where leaving just got cheaper.

Tags: Anthropic, IPO, AI economy, AT&T, open-weight AI models, Tinder